Dropout

Overview

Puts the packet-loss-concealment algorithms from SPIRAL's research directly in a plug-in. Dropout drops packets of your audio on purpose and lets one of six concealment methods — from classic autoregressive prediction to self-attention and modal resonance — fill each gap, run past the point of transparency until the “repair” becomes the effect.

How it works

Dropout simulates a network dropping packets of your audio, then asks a concealment algorithm to fill each gap. Concealment is meant to be inaudible; here it is driven hard enough that the repair becomes the sound.

Losing packets

Audio is cut into packets of 2.5 to 60 ms, the same frame sizes speech and VoIP codecs use. Density sets the long-run share of packets lost, Burst the average number lost in a row, and density can be modulated or locked to the host tempo.

Six ways to fill the gap

Autoregressive

Classic VoIP concealment. On every good packet an all-pole model is fitted to the audio by Levinson–Durbin recursion and smoothed across packets; on a lost one the model runs forward from the last samples, extrapolating the waveform.

Waveform Morph

Telephony-style waveform substitution. It finds the most recent pitch-period loop by normalised cross-correlation and replays it with two slightly detuned read heads, so the repeat slowly phases against itself.

Self-Attention

Scores several past passages against the audio just before the drop and blends what followed each, weighted by a softmax of their similarity: query/key/value attention in miniature. Low focus gives a smeared, chorused texture; high focus approaches a clean repeat.

Audio Inpainting

Matching pursuit fits the audio before the drop as a handful of sinusoids from a fixed dictionary, then simply lets them keep rotating through the gap. Few atoms sound like isolated tones; many rebuild a fuller, glassy spectrum.

DRED-Style

After Opus's Deep Redundancy. The signal is delayed while a deliberately degraded copy — decimated and bit-reduced — is stored alongside it; a lost packet is rebuilt from that lossy copy rather than predicted.

Prony Modal

Prony's method fits the audio as a sum of damped resonant modes and lets them ring on through the gap, so plucked or struck sounds seem to continue their own natural decay. A damping control rescales every mode live.